append_chunk: do not make a copy if the whole chunk is appended.

[fredrikekre]

With

function perf(bytes)
    buf = BufferStream()
    @sync begin
        @async begin
            for i in 1:1024
                write(buf, bytes)
            end
            close(buf)
        end
        @async begin
            write(devnull, buf)
        end
    end
end

This goes from

julia> @btime perf($(rand(UInt8, 1024)));
  810.652 μs (19497 allocations: 2.69 MiB)

julia> @btime perf($(rand(UInt8, 1024^2)));
  349.328 ms (21545 allocations: 2.00 GiB)

to

julia> @btime perf($(rand(UInt8, 1024)));
  644.978 μs (17449 allocations: 1.52 MiB)

julia> @btime perf($(rand(UInt8, 1024^2)));
  119.063 ms (18473 allocations: 1.00 GiB)

[fredrikekre]

With

diff --git a/src/BufferStream.jl b/src/BufferStream.jl
index e6ebdf4..546ab0e 100644
--- a/src/BufferStream.jl
+++ b/src/BufferStream.jl
@@ -86,7 +86,7 @@ end
 # Helper methods for read/write
 write(bs::BufferStream, x::UInt8) = write(bs, [x])
 function unsafe_write(bs::BufferStream, ref::Ptr{UInt8}, nbytes::UInt)
-    return append_chunk(bs, copy(unsafe_wrap(Array, ref, (nbytes,))))
+    return append_chunk(bs, unsafe_wrap(Array, ref, (nbytes,)))
 end
 
 # Read a single byte.

I get

julia> @btime perf($(rand(UInt8, 1024^2)));
  532.917 μs (16425 allocations: 386.08 KiB)

julia> @btime perf($(rand(UInt8, 1024^2)));
  532.686 μs (16425 allocations: 386.08 KiB)

but that seems too good to be true? Writing this to file and verifying the bytes give the correct thing though.

[fredrikekre]

Perhaps writing to /dev/null is not a very good benchmark. Writing to a file instead:

function perf(bytes)
    buf = BufferStream()
    @sync begin
        @async begin
            for i in 1:1024
                write(buf, bytes)
            end
            close(buf)
        end
        @async begin
            open("rand.txt", "w") do io
                write(io, buf)
            end
        end
    end
end

seems to not show a big difference since the hardware is probably limiting here.

Still, with the diff in the previous comment:

julia> @btime perf($(rand(UInt8, 1024)));
  1.279 ms (16438 allocations: 386.81 KiB)

julia> @btime perf($(rand(UInt8, 1024^2)));
  2.260 s (16438 allocations: 386.81 KiB)

is pretty nice since basically nothing is allocated

[codecov]

Codecov Report

Merging #7 (92fa5bc) into master (e592419) will increase coverage by 0.09%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master       #7      +/-   ##
==========================================
+ Coverage   95.55%   95.65%   +0.09%     
==========================================
  Files           2        2              
  Lines          90       92       +2     
==========================================
+ Hits           86       88       +2     
  Misses          4        4              

Impacted Files	Coverage Δ
src/BufferStream.jl	95.60% <100.00%> (+0.09%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update e592419...92fa5bc. Read the comment docs.